Enterprise AGI: OpenAI promises 'Astra' while 78% of LatAm boards don't govern AI

By Zendoric · September 5, 2026.

Sam Altman has said that OpenAI will release, before the end of 2026, a system the company considers AGI (artificial general intelligence, the ability to reason at a human level across multiple domains), code-named Astra and capable —according to his own description, reported by Geeky Gadgets— of solving open mathematical problems and conducting independent research faster than a human team. It's a powerful headline. But IBM CEO Arvind Krishna has said almost the opposite: the odds that today's large language models lead directly to AGI are "close to zero." Both statements can be true at the same time because they describe different things: a product promise and an assessment of the technical architecture.

For anyone running a company, that debate is secondary. What's urgent, argues analyst Dave Vellante (co-founder of SiliconANGLE), is something else: the general intelligence of a frontier model will be a commodity, what he calls "data communism" —everyone accessing the same packaged capability, with no differentiation. Competitive advantage doesn't lie in which model you contract, but in whether you turn your organization's data, processes and tacit knowledge into a governed asset that agents can reason over. That's the concept Vellante dubs "Enterprise AGI" ("data capitalism"), and one that companies such as Databricks are already packaging into a product: its Genie Ontology semantic layer learns from query logs, metric definitions and user corrections to build a map of how the business actually operates.

Latin America is the laboratory where that thesis can be measured with data. The 'State of AI in Latin America' report, from the fund Hi Ventures (led by Jimena Pardo and Federico Antoni) and covered by El Español, puts regional adoption at 71% of corporations with AI already integrated into some team or function; another 23% use it only for individual productivity, and barely 5% have yet to explore it. The most interesting figure is the gap between startups and large corporations: 53% of startups already have agents in production, compared with 30% of large corporations, and 99% of those startups use AI internally (78% have built it into their core product). The contextual detail is no small matter: 62% of those startups have ten employees or fewer. In the same report's AI readiness index, Chile leads with 72.5 points, followed by Argentina (71), Brazil (70.2), Mexico (67.5) and Colombia (53.1).

Oracle's senior director of AI and Cloud, Sandra Daza, summed up the regional paradox at the El Español panel that accompanied the report: "Latin America is outpacing the United States in the consumption of certain services enabled by language models (...) the barrier is no longer the technology." It's a fair observation about adoption, though it's worth not confusing "consumption" with "installed capacity": using a frontier model via API is cheap and accessible from a mobile phone; building the governed data asset Vellante describes demands investment and engineering that most of those same startups have yet to undertake.

And that is where the real issue emerges, more relevant than the AGI label: the speed of adoption has no counterpart in governance. According to Diligent's Board Index 2026 —cited by Infobae and by KPMG at the ASCLA gathering in Lima— barely 22% of boards have a formal AI governance policy. Put another way: the remaining 78% are deploying agents that already act on customer data or make decisions with a degree of autonomy, without a formal policy defining who answers when something goes wrong.

That autonomy, not the model's intelligence, is the real inflection point, argues Dave Link, CEO of ScienceLogic, in Forbes: the question that separates those who scale AI from those who stall at pilots isn't how smart the model is, but how much real decision-making is delegated to it within critical systems. Satya Nadella holds a similar view: success will be measured in improved workflows, not in reasoning milestones. The region's own adoption data bears this out: among large corporations, 40% maintain direct human oversight and only 5% operate fully autonomously; the rest sit in a middle ground of bounded autonomy.

There is a concrete criterion for deciding where to cede that autonomy, and it came from a case in the region itself: Iván Herrero Bartolomé, Chief Data Officer of Grupo Intercorp, presented internal company data at the ASCLA panel —in repetitive processes of low or medium complexity, AI recorded an error rate of between 5% and 8%, compared with 10%-12% for the same processes carried out by people. It's a replicable method, and more useful than any debate over whether a model "is or isn't" AGI: measure error comparatively, process by process, before automating.

This connects with something we have been repeating at Zendoric: the "AGI" label works above all as a corporate storytelling tool —for raising capital and generating headlines— rather than as a verifiable technical milestone, and it is always worth separating demonstrated capability from product promise. What is verifiable, and what really decides who wins the next decade, is more boring and more important: who governs their own data, who measures error instead of assuming it, and who assigns a human owner to every system that acts on its own. That, and not Astra's launch date, is where real competitive advantage is decided.

In the short term, the governance vacuum portrayed by Diligent's Board Index is a genuine risk, not a nuance: the more agents operating on sensitive data without a formal policy behind them, the larger the attack surface, and a security breach in that environment not only compromises data, it can also manipulate decisions that are already automated. Startups —smaller, without decades of legacy systems— have a real structural advantage here: they can design their processes assuming agents from day one, something no large corporation can replicate overnight. In the long run, that same pattern —a micro-entrepreneur with a mobile phone who today has computing power comparable to that of a large corporation five years ago, in the words of Paqui Casanueva (chair of the board of Endeavor)— is exactly the kind of cost compression that underpins the abundance thesis: when building and governing intelligence no longer requires massive capital, advantage shifts from the size of the company to the quality of its decisions. The condition, as we keep insisting, is that governance does not fall permanently behind adoption.